aboutsummaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll3176
1 files changed, 1630 insertions, 1546 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
index 3aaf254..9041f64 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
@@ -3044,91 +3044,66 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v10i32_to_v40i8:
@@ -5025,39 +5000,41 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v18.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v34.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v35.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB14_3
@@ -5071,63 +5048,53 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v15.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v22.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v23.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v12.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
@@ -5140,147 +5107,110 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v27, v9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2
; GFX11-TRUE16-NEXT: .LBB14_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v15.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v25, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v17.h, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v18.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v19.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v12.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -9991,91 +9921,66 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v10f32_to_v40i8:
@@ -11997,39 +11902,41 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v18.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v34.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v35.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB34_3
@@ -12043,63 +11950,53 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB34_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v15.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v22.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v23.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v12.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
@@ -12112,147 +12009,110 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v27, v9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB34_2
; GFX11-TRUE16-NEXT: .LBB34_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v15.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v25, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v17.h, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v18.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v19.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v12.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -16367,91 +16227,66 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v20i16_to_v40i8:
@@ -19727,212 +19562,421 @@ define inreg <20 x i16> @bitcast_v40i8_to_v20i16_scalar(<40 x i8> inreg %a, i32
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT: s_branch .LBB51_2
;
-; GFX11-LABEL: bitcast_v40i8_to_v20i16_scalar:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
-; GFX11-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
-; GFX11-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
-; GFX11-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 8, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 8, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 8, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-NEXT: s_mov_b32 s4, 0
-; GFX11-NEXT: s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB51_4
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_and_b32 s5, s0, 0xff
-; GFX11-NEXT: s_lshl_b32 s6, s1, 8
-; GFX11-NEXT: s_and_b32 s7, s2, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s3, 8
-; GFX11-NEXT: s_or_b32 s5, s5, s6
-; GFX11-NEXT: s_or_b32 s6, s7, s8
-; GFX11-NEXT: s_and_b32 s7, s16, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s17, 8
-; GFX11-NEXT: s_and_b32 s9, s18, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s19, 8
-; GFX11-NEXT: s_or_b32 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-NEXT: s_and_b32 s9, s22, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s23, 8
-; GFX11-NEXT: s_or_b32 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_and_b32 s9, s24, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s25, 8
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v23
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_and_b32 s9, s26, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s27, 8
-; GFX11-NEXT: s_and_b32 s11, s28, 0xff
-; GFX11-NEXT: s_lshl_b32 s12, s29, 8
-; GFX11-NEXT: s_or_b32 s9, s9, s10
-; GFX11-NEXT: s_or_b32 s10, s11, s12
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v22
-; GFX11-NEXT: v_and_b32_e64 v2, 0xffff, s10
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v24
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v28
-; GFX11-NEXT: v_lshl_or_b32 v4, v0, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v31
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v34
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v27
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v33
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v9, v6, v35
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v36
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v37
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v19
-; GFX11-NEXT: v_or_b32_e32 v12, v6, v17
-; GFX11-NEXT: v_lshl_or_b32 v6, v0, 16, v3
-; GFX11-NEXT: v_mov_b32_e32 v0, s5
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v32
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v5
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v21
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v8
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshl_or_b32 v7, v9, 16, v11
-; GFX11-NEXT: v_lshl_or_b32 v8, v12, 16, v13
-; GFX11-NEXT: v_mov_b32_e32 v3, s8
-; GFX11-NEXT: v_lshl_or_b32 v9, v10, 16, v14
-; GFX11-NEXT: v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-NEXT: v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-NEXT: s_cbranch_vccnz .LBB51_3
-; GFX11-NEXT: .LBB51_2: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v27
-; GFX11-NEXT: s_add_i32 s28, s28, 3
-; GFX11-NEXT: s_lshl_b32 s5, s29, 8
-; GFX11-NEXT: s_and_b32 s4, s28, 0xff
-; GFX11-NEXT: s_add_i32 s24, s24, 3
-; GFX11-NEXT: s_or_b32 s4, s5, s4
-; GFX11-NEXT: s_and_b32 s5, s24, 0xff
-; GFX11-NEXT: s_lshl_b32 s6, s25, 8
-; GFX11-NEXT: s_add_i32 s26, s26, 3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: s_or_b32 s5, s6, s5
-; GFX11-NEXT: s_and_b32 s6, s26, 0xff
-; GFX11-NEXT: s_lshl_b32 s7, s27, 8
-; GFX11-NEXT: s_add_i32 s20, s20, 3
-; GFX11-NEXT: s_or_b32 s6, s7, s6
-; GFX11-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-NEXT: s_add_i32 s22, s22, 3
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v28
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v26
-; GFX11-NEXT: s_or_b32 s7, s8, s7
-; GFX11-NEXT: s_and_b32 s8, s22, 0xff
-; GFX11-NEXT: s_lshl_b32 s9, s23, 8
-; GFX11-NEXT: s_add_i32 s16, s16, 3
-; GFX11-NEXT: s_or_b32 s8, s9, s8
-; GFX11-NEXT: s_and_b32 s9, s16, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s17, 8
-; GFX11-NEXT: s_add_i32 s18, s18, 3
-; GFX11-NEXT: s_add_i32 s0, s0, 3
-; GFX11-NEXT: s_add_i32 s2, s2, 3
-; GFX11-NEXT: s_or_b32 s9, s10, s9
-; GFX11-NEXT: s_and_b32 s10, s18, 0xff
-; GFX11-NEXT: s_lshl_b32 s11, s19, 8
-; GFX11-NEXT: s_and_b32 s0, s0, 0xff
-; GFX11-NEXT: s_lshl_b32 s1, s1, 8
-; GFX11-NEXT: s_and_b32 s2, s2, 0xff
-; GFX11-NEXT: s_lshl_b32 s3, s3, 8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v7
-; GFX11-NEXT: s_or_b32 s10, s11, s10
-; GFX11-NEXT: s_or_b32 s0, s1, s0
-; GFX11-NEXT: s_or_b32 s1, s3, s2
-; GFX11-NEXT: s_addk_i32 s5, 0x300
-; GFX11-NEXT: s_addk_i32 s6, 0x300
-; GFX11-NEXT: s_addk_i32 s9, 0x300
-; GFX11-NEXT: s_addk_i32 s10, 0x300
-; GFX11-NEXT: s_addk_i32 s0, 0x300
-; GFX11-NEXT: s_addk_i32 s1, 0x300
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v29
-; GFX11-NEXT: v_or_b32_e32 v5, v35, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v33, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v23
-; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-NEXT: s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v18
-; GFX11-NEXT: s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v25
-; GFX11-NEXT: s_addk_i32 s7, 0x300
-; GFX11-NEXT: s_addk_i32 s8, 0x300
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v20
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-NEXT: v_or_b32_e32 v5, v34, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v10
-; GFX11-NEXT: s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v16
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v0, v19, v0
-; GFX11-NEXT: v_or_b32_e32 v2, v37, v2
-; GFX11-NEXT: v_or_b32_e32 v7, v32, v7
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v22, v4
-; GFX11-NEXT: s_addk_i32 s4, 0x300
-; GFX11-NEXT: v_or_b32_e32 v1, v21, v1
-; GFX11-NEXT: v_or_b32_e32 v3, v17, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v31, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-NEXT: v_and_b32_e64 v10, 0xffff, s4
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshl_or_b32 v4, v4, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v5
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshl_or_b32 v5, v6, 16, v7
-; GFX11-NEXT: v_lshl_or_b32 v6, v11, 16, v10
-; GFX11-NEXT: v_lshl_or_b32 v7, v8, 16, v9
-; GFX11-NEXT: v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-NEXT: v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: .LBB51_3: ; %end
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB51_4:
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_branch .LBB51_2
+; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v20i16_scalar:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-TRUE16-NEXT: s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB51_4
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s17, 8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-TRUE16-NEXT: s_and_b32 s6, s16, 0xff
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s18, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s19, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s6, s7
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s20, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s21, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s23, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s6, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s8, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s25, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s26, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s27, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s28, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s29, 8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s9 :: v_dual_and_b32 v1, 0xff, v30
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v31
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v26
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v2, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v27
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v1, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v20
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, s6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v22
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v34
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v28
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v35
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, s5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v33
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v25
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v3, v37
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB51_3
+; GFX11-TRUE16-NEXT: .LBB51_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-TRUE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-TRUE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-TRUE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-TRUE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v25
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v26
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v37, v2
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v17, v3
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 0x300, v2
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v36, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v24
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v30
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v33, v4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xff, v6
+; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-TRUE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v10
+; GFX11-TRUE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v20
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v18
+; GFX11-TRUE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v34, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v19, v0
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v32, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v10.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v22, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v35, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, s1
+; GFX11-TRUE16-NEXT: .LBB51_3: ; %end
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-NEXT: .LBB51_4:
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-TRUE16-NEXT: s_branch .LBB51_2
+;
+; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v20i16_scalar:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-FAKE16-NEXT: s_mov_b32 s4, 0
+; GFX11-FAKE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB51_4
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s17, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s19, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s23, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s25, 8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v23
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s27, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s11, s28, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s12, s29, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s9, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s12
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v22
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v2, 0xffff, s10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v30
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v24
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v28
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v0, 16, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v26
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v3, v31
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v5, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v29
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v27
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v25
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v6, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v36
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v37
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v19
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v6, v17
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v0, 16, v3
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, s5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v9, 16, v11
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v12, 16, v13
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v10, 16, v14
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB51_3
+; GFX11-FAKE16-NEXT: .LBB51_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v27
+; GFX11-FAKE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-FAKE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-FAKE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-FAKE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v36, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v26
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-FAKE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-FAKE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v24
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v7
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-FAKE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-FAKE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v35, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v33, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v23
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v18
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v25
+; GFX11-FAKE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 0x300, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v34, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v30
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v10
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v19, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v37, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v32, v7
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v22, v4
+; GFX11-FAKE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v21, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v17, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v31, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v10, 0xffff, s4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v6, 16, v7
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v11, 16, v10
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v8, 16, v9
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v3, 16, v2
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-FAKE16-NEXT: .LBB51_3: ; %end
+; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-FAKE16-NEXT: .LBB51_4:
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-FAKE16-NEXT: s_branch .LBB51_2
%cmp = icmp eq i32 %b, 0
br i1 %cmp, label %cmp.true, label %cmp.false
@@ -22484,91 +22528,66 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v20f16_to_v40i8:
@@ -25790,212 +25809,421 @@ define inreg <20 x half> @bitcast_v40i8_to_v20f16_scalar(<40 x i8> inreg %a, i32
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT: s_branch .LBB63_2
;
-; GFX11-LABEL: bitcast_v40i8_to_v20f16_scalar:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
-; GFX11-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
-; GFX11-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
-; GFX11-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 8, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 8, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 8, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-NEXT: s_mov_b32 s4, 0
-; GFX11-NEXT: s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB63_4
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_and_b32 s5, s0, 0xff
-; GFX11-NEXT: s_lshl_b32 s6, s1, 8
-; GFX11-NEXT: s_and_b32 s7, s2, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s3, 8
-; GFX11-NEXT: s_or_b32 s5, s5, s6
-; GFX11-NEXT: s_or_b32 s6, s7, s8
-; GFX11-NEXT: s_and_b32 s7, s16, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s17, 8
-; GFX11-NEXT: s_and_b32 s9, s18, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s19, 8
-; GFX11-NEXT: s_or_b32 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-NEXT: s_and_b32 s9, s22, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s23, 8
-; GFX11-NEXT: s_or_b32 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_and_b32 s9, s24, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s25, 8
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v23
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-NEXT: s_or_b32 s8, s9, s10
-; GFX11-NEXT: s_and_b32 s9, s26, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s27, 8
-; GFX11-NEXT: s_and_b32 s11, s28, 0xff
-; GFX11-NEXT: s_lshl_b32 s12, s29, 8
-; GFX11-NEXT: s_or_b32 s9, s9, s10
-; GFX11-NEXT: s_or_b32 s10, s11, s12
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v22
-; GFX11-NEXT: v_and_b32_e64 v2, 0xffff, s10
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v24
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v28
-; GFX11-NEXT: v_lshl_or_b32 v4, v0, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v31
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v34
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v27
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v33
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v9, v6, v35
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v36
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v37
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v19
-; GFX11-NEXT: v_or_b32_e32 v12, v6, v17
-; GFX11-NEXT: v_lshl_or_b32 v6, v0, 16, v3
-; GFX11-NEXT: v_mov_b32_e32 v0, s5
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v32
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v5
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v21
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v8
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshl_or_b32 v7, v9, 16, v11
-; GFX11-NEXT: v_lshl_or_b32 v8, v12, 16, v13
-; GFX11-NEXT: v_mov_b32_e32 v3, s8
-; GFX11-NEXT: v_lshl_or_b32 v9, v10, 16, v14
-; GFX11-NEXT: v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-NEXT: v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-NEXT: s_cbranch_vccnz .LBB63_3
-; GFX11-NEXT: .LBB63_2: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v27
-; GFX11-NEXT: s_add_i32 s28, s28, 3
-; GFX11-NEXT: s_lshl_b32 s5, s29, 8
-; GFX11-NEXT: s_and_b32 s4, s28, 0xff
-; GFX11-NEXT: s_add_i32 s24, s24, 3
-; GFX11-NEXT: s_or_b32 s4, s5, s4
-; GFX11-NEXT: s_and_b32 s5, s24, 0xff
-; GFX11-NEXT: s_lshl_b32 s6, s25, 8
-; GFX11-NEXT: s_add_i32 s26, s26, 3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: s_or_b32 s5, s6, s5
-; GFX11-NEXT: s_and_b32 s6, s26, 0xff
-; GFX11-NEXT: s_lshl_b32 s7, s27, 8
-; GFX11-NEXT: s_add_i32 s20, s20, 3
-; GFX11-NEXT: s_or_b32 s6, s7, s6
-; GFX11-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-NEXT: s_add_i32 s22, s22, 3
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v28
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v26
-; GFX11-NEXT: s_or_b32 s7, s8, s7
-; GFX11-NEXT: s_and_b32 s8, s22, 0xff
-; GFX11-NEXT: s_lshl_b32 s9, s23, 8
-; GFX11-NEXT: s_add_i32 s16, s16, 3
-; GFX11-NEXT: s_or_b32 s8, s9, s8
-; GFX11-NEXT: s_and_b32 s9, s16, 0xff
-; GFX11-NEXT: s_lshl_b32 s10, s17, 8
-; GFX11-NEXT: s_add_i32 s18, s18, 3
-; GFX11-NEXT: s_add_i32 s0, s0, 3
-; GFX11-NEXT: s_add_i32 s2, s2, 3
-; GFX11-NEXT: s_or_b32 s9, s10, s9
-; GFX11-NEXT: s_and_b32 s10, s18, 0xff
-; GFX11-NEXT: s_lshl_b32 s11, s19, 8
-; GFX11-NEXT: s_and_b32 s0, s0, 0xff
-; GFX11-NEXT: s_lshl_b32 s1, s1, 8
-; GFX11-NEXT: s_and_b32 s2, s2, 0xff
-; GFX11-NEXT: s_lshl_b32 s3, s3, 8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v7
-; GFX11-NEXT: s_or_b32 s10, s11, s10
-; GFX11-NEXT: s_or_b32 s0, s1, s0
-; GFX11-NEXT: s_or_b32 s1, s3, s2
-; GFX11-NEXT: s_addk_i32 s5, 0x300
-; GFX11-NEXT: s_addk_i32 s6, 0x300
-; GFX11-NEXT: s_addk_i32 s9, 0x300
-; GFX11-NEXT: s_addk_i32 s10, 0x300
-; GFX11-NEXT: s_addk_i32 s0, 0x300
-; GFX11-NEXT: s_addk_i32 s1, 0x300
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v29
-; GFX11-NEXT: v_or_b32_e32 v5, v35, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v33, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v23
-; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-NEXT: s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v18
-; GFX11-NEXT: s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v25
-; GFX11-NEXT: s_addk_i32 s7, 0x300
-; GFX11-NEXT: s_addk_i32 s8, 0x300
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v20
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-NEXT: v_or_b32_e32 v5, v34, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v10
-; GFX11-NEXT: s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v16
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v0, v19, v0
-; GFX11-NEXT: v_or_b32_e32 v2, v37, v2
-; GFX11-NEXT: v_or_b32_e32 v7, v32, v7
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v22, v4
-; GFX11-NEXT: s_addk_i32 s4, 0x300
-; GFX11-NEXT: v_or_b32_e32 v1, v21, v1
-; GFX11-NEXT: v_or_b32_e32 v3, v17, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v31, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-NEXT: v_and_b32_e64 v10, 0xffff, s4
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshl_or_b32 v4, v4, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v5
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshl_or_b32 v5, v6, 16, v7
-; GFX11-NEXT: v_lshl_or_b32 v6, v11, 16, v10
-; GFX11-NEXT: v_lshl_or_b32 v7, v8, 16, v9
-; GFX11-NEXT: v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-NEXT: v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: .LBB63_3: ; %end
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB63_4:
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_branch .LBB63_2
+; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v20f16_scalar:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-TRUE16-NEXT: s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB63_4
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s17, 8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-TRUE16-NEXT: s_and_b32 s6, s16, 0xff
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s18, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s19, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s6, s7
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s20, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s21, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s23, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s6, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s8, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s25, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s26, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s27, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s28, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s29, 8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s9 :: v_dual_and_b32 v1, 0xff, v30
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v31
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v26
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v2, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v27
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v1, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v20
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, s6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v22
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v34
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v28
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v35
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, s5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v33
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v25
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v3, v37
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB63_3
+; GFX11-TRUE16-NEXT: .LBB63_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-TRUE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-TRUE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-TRUE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-TRUE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v25
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v26
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v37, v2
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v17, v3
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 0x300, v2
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v36, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v24
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v30
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v33, v4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xff, v6
+; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-TRUE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v10
+; GFX11-TRUE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v20
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v18
+; GFX11-TRUE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v34, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v19, v0
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v32, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v10.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v22, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v35, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, s1
+; GFX11-TRUE16-NEXT: .LBB63_3: ; %end
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-NEXT: .LBB63_4:
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-TRUE16-NEXT: s_branch .LBB63_2
+;
+; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v20f16_scalar:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-FAKE16-NEXT: s_mov_b32 s4, 0
+; GFX11-FAKE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB63_4
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s17, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s19, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s23, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s25, 8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v23
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s27, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s11, s28, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s12, s29, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s9, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s12
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v22
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v2, 0xffff, s10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v30
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v24
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v28
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v0, 16, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v26
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v3, v31
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v5, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v29
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v27
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v25
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v6, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v36
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v37
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v19
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v6, v17
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v0, 16, v3
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, s5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v9, 16, v11
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v12, 16, v13
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v10, 16, v14
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB63_3
+; GFX11-FAKE16-NEXT: .LBB63_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v27
+; GFX11-FAKE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-FAKE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-FAKE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-FAKE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v36, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v26
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-FAKE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-FAKE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v24
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v7
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-FAKE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-FAKE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v35, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v33, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v23
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v18
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v25
+; GFX11-FAKE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 0x300, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v34, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v30
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v10
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v19, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v37, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v32, v7
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v22, v4
+; GFX11-FAKE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v21, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v17, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v31, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v10, 0xffff, s4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v6, 16, v7
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v11, 16, v10
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v8, 16, v9
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v3, 16, v2
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-FAKE16-NEXT: .LBB63_3: ; %end
+; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-FAKE16-NEXT: .LBB63_4:
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-FAKE16-NEXT: s_branch .LBB63_2
%cmp = icmp eq i32 %b, 0
br i1 %cmp, label %cmp.true, label %cmp.false
@@ -28791,39 +29019,38 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v37.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB72_3
@@ -28837,65 +29064,55 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB72_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v30.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v26.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
@@ -28906,146 +29123,110 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB72_2
; GFX11-TRUE16-NEXT: .LBB72_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v23.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v22.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v21.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v24.h, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v25.h, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v26.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v27.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v16.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v18.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -30878,91 +31059,66 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v5f64_to_v40i8:
@@ -32912,39 +33068,38 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v37.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB76_3
@@ -32958,65 +33113,55 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB76_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v30.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v26.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
@@ -33027,146 +33172,110 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB76_2
; GFX11-TRUE16-NEXT: .LBB76_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v23.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v22.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v21.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v24.h, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v25.h, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v26.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v27.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v16.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v18.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -35022,91 +35131,66 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l
; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v5i64_to_v40i8: